Weakly Supervised Social-Satellite Fusion for Near Real-Time Soil Erosion Detection and Linguistic Analysis in the United States

Soil erosion poses a pervasive threat to agricultural sustainability and infrastructure resilience
across the United States, yet conventional monitoring, reliant on sporadic field surveys or costly,
expert-driven image analysis, fails to deliver the spatiotemporal resolution necessary for
proactive management. To overcome these limitations, this study introduces a scalable, near real
time framework that synergizes crowdsourced social media data with open-access satellite
imagery through a weakly supervised deep learning approach. We curated a dataset of over 500
geotagged tweets (2015–2024) containing erosion-related keywords (e.g., “mudslide,” “washed
out road”), which served as noisy yet geographically rich labels for training a Convolutional
Neural Network (CNN) on Google Colab. The model was trained on corresponding 10–30 m
image patches from Sentinel-2 and Landsat 8/9, processed via Google Earth Engine. To ensure
spectral consistency across sensors and minimize data heterogeneity, we implemented a rigorous
harmonization workflow: (1) bidirectional reflectance distribution function (BRDF) correction
and atmospheric normalization using the Sen2Cor and LaSRC algorithms, respectively; (2)
spectral band-pass adjustment to align Landsat and Sentinel-2 surface reflectance into a common
radiometric framework; and (3) temporal aggregation into monthly cloud-free composites via
median mosaicking to mitigate data gaps. From these harmonized stacks, we derived indices
including NDVI and NDWI, enabling the CNN to predict erosion severity with an F1-score of
0.82 against the USDA’s SSURGO database, validating the framework's efficacy in identifying
high-risk zones. Concurrently, Natural Language Processing (NLP) of the tweet corpus
unearthed regional terminological fingerprints (e.g., “gully” in the Midwest vs. “washout” in
Appalachia), which strongly correlated with distinct geomorphic processes and severity levels.
This work is the first to operationalize social media as a dynamic "social sensor" for soil loss,
offering a low-cost, field-deployable alternative to traditional methods. The integrated
framework democratizes access to timely erosion intelligence, providing actionable insights for
disaster response, targeted land-use policy, and community resilience planning.